Structured Data Leads to Better Analysis, Outcomes

Software that can create structured tables of data from clinicians' notes and then incorporate them into any standard electronic medical record, is easing concerns that structured EHRs are killing the clinical narrative.

At CHIME earlier this month, I heard many CIOs complain that electronic health record systems do a poor job of summarizing clinicians' notes and integrating them with the structured data which forms the backbone for much of the population health analytics which can bend the cost curve of care.

I've been writing for a long time about concerns that structured EHRs are abandoning the clinical narrative. I've even written about the potential for natural language processing (NLP) technology to extract actionable information from that narrative.

Now there is evidence that NLP is starting to make a difference, and more importantly, may not require providers to be locked into a new set of such technologies. Instead, providers might be able to shop around for best-of-breed tools to get the job done.

The reason for my optimism is IBM's LanguageWare Content Analytics software, now in use at the University of North Carolina Health Center.

IBM's software can actually create structured tables of data from free text, which can then be incorporated into any standard medical record, according to IBM officials.

At UNCHC, the software is digging into written mammography reports and finding abnormal results, then presenting them for followup examinations, says Carlton Moore, associate professor of medicine at UNCHC.

"We looked at a random sample of mammography reports taken from our electronic medical records done over the past five years," Moore says. Two physicians reviewed the reports; then IBM's software went through the same reports, and the team compared the two findings.

The software found 98 percent of the abnormalities that the physicians found. "It was actually very effective" and could be tweaked to be 100 percent effective, Moore says. UNCHC's results have been written up and submitted to a research journal for possible publication.

At present, UNCHC has a home-grown electronic medical record, but is in the process of switching over to Epic by next May, and is looking to integrate IBM's NLP software with Epic after that.

Moore says physicians' crazy day-to-day workflow makes a place for NLP to flag abnormalities for followup that otherwise would be overlooked.

"There's a lot of information coming physicians' way and they have to process it," Moore says. "They have a lot of interruptions. They're writing a note and might get interrupted, because a patient just called or a nurse wants you to see a patient right away, so it's very easy for things to kind of fall through the cracks."

Physicians should not have to rely upon memory to go back and make sure that an abnormality is followed up on, Moore says.

Providers will also be given options on how to follow up. In some cases, this analysis can trigger an alert in an EHR. In others, a daily report could be routed to nurse care managers to look through the report and make sure all patients have had proper follow up care, Moore says.

The same NLP techniques UNCHC is using to scrutinize mammography reports and pathology reports could also be used to scan other kinds of radiology reports, or any other type of free text reports.

By taking the next step and converting its results into structured tables that can be incorporated into patient EHRs, UNCHC and other IBM customers will avoid being locked into some proprietary NLP system that would sit alongside the traditional EHR.

That's not to say IBM doesn't want customers to stay with its solution. But by getting into that structured format, it does tilt the balance of power back toward the customer, because structured data is inherently more interoperable and portable than unstructured data. As IBM's Ed Macko, worldwide CTO for healthcare and life sciences puts it, the newly-structured data, extracted by NLP, becomes part of the patient's longitudinal record.

You might not think it takes a rocket scientist to figure this out, but you might be wrong. Prior to IBM, Macko, an engineer by training, built computer systems for the space shuttle.

Seriously, NLP can uncover the hidden truths in the patient record. Take the elusive condition known as smoking status. A clinician could check the box indicating the patient does not smoke. But the clinical narrative reveals a more nuanced reality. Macko notes that a physician's note may say that the same patient is "down to two packs a day." Only technology such as NLP, or an army of people scanning physicians' notes, could possibly get to the single version of that truth.

There may even be new business opportunities for larger systems such as UNCHC to align with smaller healthcare systems and hospitals who can't afford the resources to implement IBM's LanguageWare Content Analytics.

Here's how it might work. Small providers could connect to systems such as UNCHC via health information exchanges. UNC's own HIE could provide NLP services to such providers, who would send their free-text reports and receive structured data and actionable reports in return, for a fee.

"That's where I see the future of this is going," Moore says. "For a small practice to be able to independently take the software and have a programmer do this themselves is probably not going to happen."

There's also a role for this kind of service in research. Many inclusion and exclusion criteria important to clinical trials are buried in physicians' notes. NLP could be used to identify these patients for use in study cohorts, Moore notes.

One other note: Although you might hear IBM describing NLP as an element of Watson, it is but one element of Watson. While Watson does employ NLP, it also contains a machine-learning element that helps Watson understand the entire ontology of a particular medical practice.

Over time, Watson "learns" more than it used to about a given practice, such as oncology. The NLP technology in use at UNCHC must be specifically programmed to seek and extract unstructured data into structured data, and as such, falls far short of Watson's ability to understand medicine.

Still, NLP is probably our best tool for mining unstructured data, and arrives none too soon given the explosion of electronically-stored medical data. And as Macko notes, "in the end, physicians are going to do what physicians do. They like to write things down, right?"